This research proposal investigates the robustness of Mamba-based State Space Models in autonomous driving systems under extreme ego-vehicle acceleration scenarios. It proposes a methodology combining theoretical analysis, empirical evaluation, and diagnostic tools to identify failure modes and enhance model stability.
Key findings
Mamba-based architectures show promise in autonomous driving due to linear computational complexity.
Proposed research identifies critical failure modes under rapid acceleration and high-frequency ego-motion variations.
Introduction of acceleration-conditioned failure taxonomy and temporal coherence validation module.
Limitations & open questions
Research is still in the proposal stage with preliminary analysis.
Further empirical validation on a broader range of scenarios is needed.